
From  Christoph Birkel <[email protected]> 
To  [email protected] 
Subject  Re: st: Re: test for time trend 
Date  Tue, 14 Feb 2006 19:55:36 +0100 
Cornelia Schmidt schrieb:
Dear Mike and Clive,The HadriLMTest (for which an adofileexists) allows to test the null of stationarity of all series. Unfortunately, according to simulation studies, its size explodes as soon as there is some serial correlation in the data, so it is quite useless in practice.
thank you very much for your suggestions. To provide a bit more information:
the dependent variable of my analysis is education spending (% GDP). I decided to limit the period under investigation to 19802000. The panel is highly unbalanced due to a large number of missings on the dependent variable.
In order to test whether there is a positive timetrend in he education spending data, I first did a Fisher test for panel data. Compared to other stationarity tests, it has the advantage of being feasible for unbalanced panels. Based on the pvalues of individual unit root tests, it assumes that all series are nonstationary under the null hypothesis against the alternative that at least one series in the panel is stationary. This reveals drawbacks of the test. In my analysis, I would actually be interested in testing a null assuming stationarity of the data. The pvalues of the Fisher test do not tell anything about that. If the null hypothesis is rejected this does not imply that all series are stationary. Therefore, in my opinion, the power of the test is quite limited.
The basic difference between the ADF and Phillips PerronTests is that the former uses a parametric correction for serial correlation (adding lagged differences), whereas the latter a nonparametric one, which is known to be very problematic (i.e.: leading to large size distortions) if there is a negative MAprocess in the residuals; so ADF might be preferable. Independent of the test you use, the power will be very low with series of this length (T=20), so it is no wonder that the unitroot hypothesis cannot be rejected.
To get a better understanding of whether the series of some countries are nonstationary, I tried to run timeseries stationaritytests for every country. And here I have some questions: I tried a DFGLS unitroot test but it seems that it does not work with missings in the time series. Alternatively, I used a PhillipsPerron and an Augmented DickeyFuller unitroot test which both worked with my data. Do you know which of these tests best fits? Can they be used alternatively or in which respects do they differ?
Including a time variable would be valid only under the assumption that there is a deterministic time trend in your data, while your results point to a stochastic trend. You might think about periodspecific fixed effects , which would take a stochastic trend common to all units out of the data (but not individualspecific unitroot processes). Alternatively, if you are interested in longrun relationships and some of the covariates are also nonstationary, you might consider testing for cointegration (unfortunately there are no such tests available in STATA, but a rough test would be to check for stationarity of the regression residuals) and estimating an ErrorCorrectionModel.
The test results indicate that I cannot reject the null hypothesis of nonstationarity for most of the countries. As I have so many missings in my dataset, it is hardly possible to use a firstdifference estimation to correct for the time trend. Do you know an alternative way? Would it be an option to take a time variable into the model?
From: Michael Hanson <[email protected]> ReplyTo: [email protected] To: [email protected] Subject: Re: st: test for time trend Date: Sat, 4 Feb 2006 19:38:40 0500 On Feb 4, 2006, at 4:29 PM, Cornelia Schmidt wrote:I know it is a basic question, but it would be great if you could help me.
I have a panel dataset for a sample of 50 countries over a period of 40 years.
How can I test if there is a time trend in the data and how can I correct for it?
Individual trends or common? Deterministic or stochastic? Linear or otherwise? You'll need to provide a bit more information to get an appropriate answer to your question. While you could simply dummy for years, that may be effectively throwing away useful information in the time domain (certainly year dummies won't help identify a time trend); with T = 40 as you have, you may want to consider an explicit model of the timeseries process(es) contributing to your data generation process. Hope that helps.
 Mike
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
_________________________________________________________________
Sie suchen EMails, Dokumente oder Fotos? Die neue MSN Suche Toolbar mit WindowsDesktopsuche liefert in sekundenschnelle Ergebnisse. Jetzt neu! http://desktop.msn.de/ Jetzt gratis downloaden!
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
© Copyright 1996–2024 StataCorp LLC  Terms of use  Privacy  Contact us  What's new  Site index 